Deep dual dynamic gated convolutional model with interpretable pooling for genomic signal prediction

Genomic signal prediction is a key component of computer-aided diagnosis, and it aims to obtain functional contextual patterns from raw signals. Many advanced algorithms overlook the crucial contextual interactions, limiting the applicability of genomic signals. In this work, a dual dynamic gated convolutional model is devised with interpretable pooling to conduct contextual information extraction from spatial and channel levels. The proposed model could enhance the robustness of genomic signal prediction via global expectation pooling and gated mechanisms. First, we design the novel spatial-gated and channel-gated modules to integrate spatial contextual representations of genomic signals adaptively. Second, an interpretable learning algorithm with probabilistic models is developed to capture the complex interaction patterns via an interpretable pooling. Moreover, the spatial gated module estimates essential spatial features via a dual dynamic transformation mechanism, and the channel-gated module detects the salient channel patterns via a gated normalization convolutional unit. By conducting extensive experiments, the prediction performance of our model surpasses the state-of-the-art models, which substantiates the exceptional efficacy in genomic signal prediction based on unsupervised k-mer embedding.